一种改进的快速深度图像先验降噪模型  

Improved accelerated deep image prior-based denoising model

作　　者：张睿[1] 程晓慧 Zhang Rui;Cheng Xiaohui(Wuxi Vocational Institute of Commerce,Wuxi Jiangsu 214000,China;School of Computer Information Engineering,Nanchang Institute of Technology,Nanchang 330044,China;School of Mathematics&Computer Sciences,Nanchang University,Nanchang 330031,China)

机构地区：[1]无锡商业职业技术学院,江苏无锡214000 [2]南昌理工学院计算机信息工程学院,南昌330044 [3]南昌大学数学与计算机学院,南昌330031

出　　处：《计算机应用研究》2025年第3期920-926,共7页Application Research of Computers

摘　　要：相较于有监督深度降噪模型,仅利用给定的噪声图像本身就能完成降噪任务的无监督深度图像先验(deep image prior,DIP)降噪模型没有数据偏向(data bias)问题,具有更好的泛化能力。然而,DIP降噪模型较长的迭代训练步数导致其在执行效率方面仍有较大提升空间。为此,提出了一种改进的增速深度图像先验降噪模型(improved accelerated deep image prior-based denoising model,IADIP)。首先,使用多个主流有监督降噪模型处理输入的噪声图像,得到多个互补的初步降噪图像(称为预处理图像)。其次,以预处理图像作为网络输入并同时将预处理图像和噪声图像共同作为目标图像以降低DIP网络映射难度,为将DIP默认的4层UNet骨干网络简化为1层结构打下基础,从而大量减少迭代更新网络参数的计算代价。最后,在IADIP无监督迭代训练中,提出一种采用下采样技术实现的伪有参考图像质量度量,并基于该度量监控迭代过程中网络输出图像的图像质量,适时终止迭代训练以实现自动早停并确保网络输出图像的图像质量。当迭代终止时,IADIP网络输出图像即为最终的降噪后图像。大量实验表明:改进后的IADIP降噪模型的执行效率显著优于原DIP降噪模型,而其降噪效果也超过了当前主流的降噪算法。Compared to supervised deep denoising models,the unsupervised deep image prior(DIP)denoising model,which only utilizes the given noisy image itself to perform denoising,does not suffer from the problem of training data bias and exhibits better generalization capability.However,the longer training iterations of the DIP denoising model still leave room for significant improvement in terms of execution efficiency.Therefore,it proposed an improved accelerated deep image prior-based denoising model(IADIP).Firstly,it used multiple mainstream supervised denoising models to process the input noisy image,resulting in multiple complementary preliminary denoised images(referred to as preprocessed images).Then,the DIP network reduced the mapping difficulty by utilizing the preprocessed images as network inputs and simultaneously utilizing both the preprocessed images and the noisy image as target images.This set the foundation for simplifying the default 4-layer UNet backbone network of DIP into a 1-layer structure,thereby significantly reducing the computational cost of iterative parameter updates.Finally,in the unsupervised iterative training of IADIP,it proposed a pseudo full-reference image quality metric based on downsampling technique to monitor the quality of network output images during the iteration process,and timely terminated iteration training to achieve automatic early stopping.When the iteration stops,the IADIP network output image becomes the final denoised image.Extensive experiments demonstrate that the improved IADIP denoising model exhibits significantly improved execution efficiency compared to the original DIP denoising model,while also surpassing the performance of current state-of-the-art denoising algorithms.

关 键 词：图像降噪 深度图像先验 性能提升 简化网络 下采样 自动早停 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]